Discovery of Sustainable Refrigerants through Physics-Informed RL Fine-Tuning of Sequence Models

TL;DR

The paper tackles the challenge of discovering sustainable refrigerants under environmental and safety constraints with limited data. It introduces RefGen, a physics-informed RL framework that couples SMILES-based sequence models with physics-grounded property predictors and full vapor-compression cycle simulations, including the Peng-Robinson EOS and NASA polynomials. The approach uses supervised fine-tuning followed by reinforcement learning with multi-property rewards and a diversity mechanism to generate de novo candidates while ensuring thermodynamic feasibility and environmental lower-GWP impact. The results demonstrate robust predictor performance and the ability to generate novel refrigerants that balance COP, Q_vol, Tc, and GWP, including non-PFAS candidates, highlighting practical pathways for accelerated refrigerant discovery and validation in real-world contexts.

Abstract

Most refrigerants currently used in air-conditioning systems, such as hydrofluorocarbons, are potent greenhouse gases and are being phased down. Large-scale molecular screening has been applied to the search for alternatives, but in practice only about 300 refrigerants are known, and only a few additional candidates have been suggested without experimental validation. This scarcity of reliable data limits the effectiveness of purely data-driven methods. We present Refgen, a generative pipeline that integrates machine learning with physics-grounded inductive biases. Alongside fine-tuning for valid molecular generation, Refgen incorporates predictive models for critical properties, equations of state, thermochemical polynomials, and full vapor compression cycle simulations. These models enable reinforcement learning fine-tuning under thermodynamic constraints, enforcing consistency and guiding discovery toward molecules that balance efficiency, safety, and environmental impact. By embedding physics into the learning process, Refgen leverages scarce data effectively and enables de novo refrigerant discovery beyond the known set of compounds.

Figures (10)

• Figure 1: The Refgen framework. The LLM is supervised on molecular corpora for valid SMILES generation. During RL fine-tuning, the grouped predictor outputs form a multi-property reward that hooks into the LLM for policy updates. Note : an in-depth pipeline schematic can be found in Appendix \ref{['appendix:detailed_pipeline']}.
• Figure 2: Vapor-compression cycle with four states. Each state is labeled with its specific enthalpy $h_i$, and state 1 also shows the specific volume $v_1$, used in defining the volumetric refrigerating effect $Q_{\text{vol}}$. The coefficient of performance (COP) follows from enthalpy differences between states.
• Figure 3: Reward weights
• Figure 4: Mode collapse with entropy maximization.
• Figure 5: Comparison between ground-truth saturation dome from CoolProp (blue) and predicted (red) for molecules Neopentane (left) and R134a (right). The COP is computed for condenser and evaporator temperatures at 10°C and 40°C respectively. Dome translation shifts are due to varying reference points in the enthalpy and they are irrelevant for the computation of COP or $Q_{vol}$.